GRoberta: A hybrid deep learning model for fake news detection

The rapid spread of fake news on digital platforms has become a critical challenge, threatening the integrity of information and influencing public opinion. The traditional detection techniques, such as manual verification and classical machine learning, are ineffective because of the constraints on processing capacity. To address this challenge, this study proposes GRoberta, a novel hybrid deep learning model integrating Roberta and GPT-۲ to enhance stance classification performance. The model retains the contextual value of Roberta and the generative characteristics of GPT-۲ by combining their last hidden states and feeding them through a dense classification layer. To conserve efficiency, the lower layers of both transformer models are kept static during training, thereby saving computational costs without compromising performance levels. The model is tested on the FNC-۱ dataset using stratified five-fold cross-validation based on an ۸۰-۱۰-۱۰ split for training, validation, and testing. GRoberta performs remarkably, with accuracy at ۹۶.۵%, precision at ۹۶.۵۲%, and recall at ۹۶.۵%, better than most tested benchmarks. Improvement in all evaluation metrics, as well as minimal residual error, testify to the reliability and effectiveness of the model. Based on similar study analysis, an improvement of ۲.۸۵% in accuracy proves that GRoberta is an effective, high-performance approach to fake news detection.

Keywords:

Fake News Detection , Hybrid Deep Learning , Social media , Large Language Model , Transformers